Tenth article, its cover, and short 14 — the add-in's divulgation

[peopleworks]

The standing rule: no big feature is done without the article, the shorts and the posts, bilingual. This is the Word add-in's half.

The angle is not the launch

It is what the pane did the first time it ran on a real document: produced a score, refused the verdict at 357 words, and reported six no-break spaces anyway. That is the project's whole argument, printed inside the application people actually write in — and it arrived as a screenshot rather than as a claim.

Article	signsofai-word-story.{en,es}.md + built HTML. ~1,600 words each, written in each language rather than translated
Cover	social/word-story-cover.svg + PNG, rasterised with headless Chrome
Short 14	ES and EN HTML + scenes, timings in fractions of --dur
Copy	X and LinkedIn, both languages, in PUBLICACION.md

Both articles measure 0/100 against our own engine

Which is the house rule for anything published. The English one started at 7/100 with five findings and I took the advice rather than waiving it: two actually as empty intensifiers, two mid-sentence bolds, and three em-dashes where a colon, a full stop and a comma read better.

The cover argues rather than decorates

A ruler to scale with 357 to the left of 649, the pane beside it with the score in grey and the verdict box drawn empty, and the U+00A0 ×6 row underneath. Every figure is real — from published-calibration.json and that first run.

Short 14, and the trap that bit

Timings in fractions of --dur so it can be narrated later, the technique shorts 12 and 13 established.

Two layout breaks that only a snapshot shows, and the second is the trap this very file documents in a comment three blocks above where I made it: the strike-through's keyframe wrote transform without the rotate, so it snapped flat the instant it animated. The English privacy line also wrapped to three lines where Spanish takes two, and walked into the footer. Both fixed and re-snapped in both languages.

Left for Pedro

Narration needs the ElevenLabs key. Narrate 14, put the measured duration into --dur, render — the command is in GUION-shorts.md, alongside 12 and 13 which are still waiting on the same thing.

🤖 Generated with Claude Code

https://claude.ai/code/session_015PEbbiYSNPw7jE3LrPNhyF

[peopleworks]

Shorts 12, 13 and 14 are narrated and rendered — six videos

The key never needed handing over: narrate-all.mjs already points at the kit, and narrate.mjs reads .elevenlabs.key from there. Done.

	ES	EN
12 — the assistant's own turn	22.1s voice → 25.0s	22.4s → 25.0s
13 — 94/100 and no verdict	23.0s → 27.0s	23.9s → 25.3s
14 — inside Word, and it goes quiet	24.0s → 26.0s	26.8s → 28.2s

The fraction technique worked exactly as designed. All three were written before the voice existed; setting one number per file put every beat where it belonged, and nothing had to be re-timed. That closes the loop on what 12 and 13 were experimenting with.

Two things now written into GUION-shorts.md

• --dur is the narration's length, not the video's. The video adds tailSeconds, and minSeconds can stretch it past both. The fractions distribute phrases inside the speech, so they belong to the speech.
• English and Spanish do not last the same, even saying the same thing: 24.0 against 26.8 on short 14. Each language needs its own --dur; one shared value desynchronises whichever is wrong.

Verified by pulling a late frame out of the rendered mp4, not just by trusting the build log.

Renders and subtitles are in out/, which is git-ignored, so this commit carries the durations and the script rather than the files.